The Life Cycle of Corporate Venture Capital

Transcription

1 The Life Cycle of Corporate Venture Capital Song Ma Abstract This paper investigates why industrial firms conduct Corporate Venture Capital (CVC) investment in entrepreneurial companies. I test alternative views on CVC by exploiting the entry, investment, and termination decisions of CVC divisions. CVC entry follows deterioration of a firm s internal innovation. At the investment stage, CVCs select startups with a similar technological focus but that have a non-overlapping knowledge base, and they integrate technologies generated from these ventures. CVCs are terminated when parent firms innovation recovers. Overall, the strategic desire to regain innovation after adverse shocks, rather than managerial misbehaviors or pure financial returns, motivates firms to adopt CVCs. JEL Classification: G24, G34, O32, D83 Keywords: Corporate Venture Capital, Innovation, Entrepreneurship, Investment, R&D This paper is based on portions of my dissertation submitted to Duke University. I am indebted to my dissertation committee: John Graham (co-chair), David Robinson (co-chair), Manuel Adelino, Alon Brav, Manju Puri, and Ronnie Chatterji. I have also benefited from discussions with Jean-Noël Barrot, Mike Ewens, Philippe Gorry, Yael Hochberg, Alan Kwan, Josh Lerner, Xiaoding Liu, William Mann, Justin Murfin, Bruce Petersen, Mark Schankerman, Felipe Severino, Morten Sørensen and Ting Xu. Seminar participants at AFA (Chicago), CKGSB, CICF (Xiamen), Copenhagen Business School, Cornell, Dartmouth PE Conference, Duke, Duke Entrepreneurship and Innovation Seminar, FIRS (Lisbon), Harvard, HKU, HKUST, Imperial College, London Business School, Maryland, Michigan State, NC State, NFA (Mont Tremblant), Ninth Annual Searle Center/USPTO Conference on Innovation Economics, Rochester, Toronto, UNC, Utah, WUSTL Corporate Finance Conference, and Yale provided valuable comments. All errors are my own. Please address correspondence to Song Ma, 165 Whitney Ave, New Haven, CT 06511; song.ma@yale.edu.

2 Recent decades have witnessed non-financial firms forays into venture capital by creating Corporate Venture Capital (CVC) divisions. That is, these firms create internal CVC divisions to make systematic minority equity investments in innovative startups. As an illustration, consider GM Ventures, 1 the CVC unit initiated by General Motors in On behalf of General Motors, GM Ventures invested in dozens of auto-related technological startups through minority equity stakes in early-stage venture rounds. The case of GM Ventures is hardly an isolated occurrence. CVC has become a common form of corporate investment adopted by hundreds of firms and has emerged as an important source of entrepreneurial capital. It accounts for about 20% of VC investment (National Venture Capital Association, 2016), and is comparable in size to 5% of US corporate research and development (R&D) spending. 2 The question naturally arises: why do firms go beyond traditional corporate investment to make arms-length entrepreneurial investments in startup ventures? Classic corporate finance theories provide several distinct, yet mutually non-exclusive, views to guide the exploration of the CVC rationale. First, the agency view of CVC would argue that the existence of CVC is rooted in the fundamental conflict between shareholders and managers. Prior literature shows that managers extract private utility by expanding firm boundaries, which in turn affects their investment decisions (Jensen, 1986; Stulz, 1990; Denis et al., 1997). CVCs, if not properly structured and monitored, might simply reflect managerial desire to build an empire or to enjoy managerial perks via venture investing (Gompers and Lerner, 2004). Alternatively, CVCs can be motivated by incumbent firms desire to earn purely financial returns from a promising entrepreneurial sector. This financial view of CVC would argue that incumbent 1 More information on this GM Ventures example can be found at 2 US Corporate R&D information is extracted from the National Science Foundation at 1

3 firms shift from internal investment to external investment such as CVC when internal investment opportunities are poor, following the classic Q-theory argument (Hayashi, 1982). Meanwhile, being affiliated with an incumbent firm allows a CVC to exploit its advantageous knowledge about the industry generated from the parent firm s core business. This advantageous knowledge is critical to mitigating information asymmetries and facilitating monitoring in funding early-stage companies (Kaplan and Strömberg, 2003). Last but not least, the strategic view of CVC emerges from several theories arguing that CVCs can be used to seek strategic complementarities from connecting to startups, most noticeably to expose firms to startups new technologies. This information acquisition function can strengthen CVC parents own internal innovation abilities. In the model of Hellmann (2002), firms make CVC investments when there are complementarities between startups and parent firms core businesses. Mathews (2006) theorizes that the main strategic benefits can be in the form of knowledge transfers from entrepreneurs to incumbent CVC parent firms. 3 Lerner (2012) argues that CVC is an important component in the architecture of corporate innovation. This paper aims to empirically investigate these different conceptual views of the CVC rationale. Understanding the CVC rationale is important for shareholders who need to govern and monitor CVC adoptions, for startups and venture capitalists who work with CVCs in entrepreneurial financing, as well as for policy makers who regulate interactions between firms and aim to stimulate innovation. To achieve this goal, I compile a comprehensive sample of CVC divisions launched by US-based public firms in the past three decades using information from both archival data and media searches. This sample is augmented by information on CVC investment history, portfolio 3 Surveys among CVC practitioners provide support for this idea, indicating that CVC investments allow parent firms to acquire information on new innovation and markets (Siegel et al., 1988; Macmillan et al., 2008). 2

4 companies, and parent firms innovation, financials, and governance. This detailed dataset allows me to empirically study each stage of the CVC life cycle from when firms enter CVC, to how CVCs invest in and interact with portfolio companies, to the decision of terminating CVCs. The key insight that helps to distinguish between the agency, financial, and strategic views is that each view generates different predictions at each stage of the CVC life cycle. The analysis begins with the CVC entry decision. I explore the determinants of CVC entry in a firm-year panel. The main finding is that typically CVCs are started following deteriorations in internal innovation, reflected in decreases in innovation quantity and quality. Quantitatively, a one-standard-deviation decline in innovation quantity, measured using the annual number of new patent applications, increases the probability that a firm will initiate a CVC in that year by about 26% relative to the unconditional entry probability. Similarly, a one-standard-deviation decline in innovation quality, measured using new patents lifetime citations, increases the entry probability by 34%. This finding supports the strategic view of CVC, which builds upon the long-held theory of information acquisition and innovation (Nelson, 1982; Telser, 1982). It argues that firms increase external information acquisition to complement internal innovation when the ability to internally generate ideas deteriorates. 4 The evidence could also be consistent with the financial view, which predicts that firms substitute internal R&D with CVC investment when internal innovation opportunities dry up. In contrast, measures of corporate governance, including institutional shareholding and the G-Index, do not explain CVC entry decisions, which indicates a lack of support for the agency view. However, one may worry that the aforementioned relation between innovation deterioration and CVC 4 See also Nelson and Winter (1982), Dosi (1988), Jovanovic and Rob (1989), Kortum (1997), Fleming and Sorenson (2004), and Frydman and Papanikolaou (2017), among others. 3

5 initiation could result from unobserved agency forces. For example, an entrenched manager could simultaneously destroy internal innovation and launch CVC as a perk, in which case the agency view would still be in play. To assess this argument, I isolate variations to innovation that are plausibly unrelated to contemporaneous managerial (mis-)behaviors. Specifically, I make use of detailed patent citation patterns to construct a Knowledge Obsolescence variable to track the usefulness of a firm s pre-determined knowledge accumulations. 5 I find that knowledge obsolescence predicts an individual firm s innovation quantity and quality declines, and the relation between these deteriorations and CVC launches is unaffected when exploiting these plausibly exogenous variations to innovation ability. This finding further rules out the agency view. The CVC entry specification is refined to further explore whether the evidence leans more toward the financial view or the strategic view. Under the financial view, entering CVC is driven by higher external Q and should be accompanied by a decrease in internal investment. However, I do not find evidence that CVC investment is accompanied by a shift away from internal R&D. If anything, internal R&D slightly increases with CVC entry, further supporting the strategic argument that CVC complements organic innovation. In another entry analysis, CVCs are classified into strategic CVCs and financial CVCs based on the corporate announcements and media coverage at the point of entry. Strategic CVCs are the majorities, and the decline of internal innovation mostly motivates entries of strategic CVCs but not financial CVCs. Overall, though the findings are not conclusive yet at this stage, the strategic view is more consistent with the empirical evidence at CVC entry. I next explore the investment phase of the CVC life cycle, hoping to further distinguish between 5 Imagine that knowledge accumulations on steam engines become less useful in the age of electric cars. Details of the variable will be provided in Section 2 and Appendix C. 4

6 the strategic and financial motives. In this stage, I examine the selection of portfolio companies, and I investigate whether and how CVC parent firms innovation paths are affected by their portfolio companies. I find that the technological proximity between the patent portfolios of a CVC parent firm and a startup has a positive effect on the probability that a venture relation will be formed. But more importantly, conditional on working in proximate technological areas, CVCs are more likely to invest in startups about which they have less information, reflected by fewer mutual citations. In addition, the prior literature demonstrates that financial return-driven, independent VCs (IVCs) exhibit a local bias when selecting portfolio companies in order to take advantage of local knowledge and to facilitate monitoring (Cumming and Dai, 2010; Hochberg and Rauh, 2012). In contrast, CVCs appear to have a reverse home bias that is, they are less likely to invest in companies in their own geographic regions, with which there are already strong local innovation spillovers (Peri, 2005; Matray, 2016). In sum, CVCs invest in companies about which they do not necessarily have advantageous knowledge but can provide more complementary innovation knowledge. This evidence lends further support to the strategic view of CVCs. Can parent firms integrate such complementarities into internal innovation? The answer is yes. CVC parent firms are more likely to cite patents generated by their portfolio startups after making the investment. This citation pattern only happens after investments are made, and never before. This pattern does not hold for placebo-pairs constructed by pairing CVC parents closely matched industry peers with startups. The integration of startups knowledge into CVC parent firms is accompanied by active adjustment of innovative human capital. Overall, CVC appears to be an important component in the system of gaining innovation. The final analysis concerns the termination stage of CVCs. 6 In principle, CVCs are not 6 The CVC life cycle is terminated when CVC parents stop making incremental investments in new startups. 5

7 constrained by the typical IVC fund life of 10 to 12 years. Surprisingly, CVCs appear to be temporary divisions that have shorter and non-uniform life cycles. The median duration of the CVC life cycle is about four years, with an average of six. I show that a CVC s staying power is closely related to the innovation dynamic of the parent firm, and it is terminated when internal innovation begins to recover. The staying power and termination decision are not explained by exit failures of portfolio companies or by governance changes such as CEO turnover. In summary, this paper investigates different views of CVCs using life cycle evidence across the entry, investment, and termination stages. The findings lend the strongest support to the strategic view CVCs are in general temporary corporate divisions for incumbent firms to respond to negative innovation shocks and help those firms to expose themselves to new technologies in order to regain their innovation edge. The agency view and the financial view, though plausible in some cases or certain stages, cannot consistently explain the large-sample life-cycle patterns. This paper contributes to an emerging literature on CVC. In prior literature, Hellmann, Lindsey, and Puri (2008) exploit a bank-vc setting and show that banks use their venture capital arms to build early relationships with startups that have larger future debt capacity, which complements their lending business. Taking CVC as given, Dushnitsky and Lenox (2005b, 2006) show that CVC investments positively correlate with parent firms future internal innovation rates and firm values, and Chemmanur, Loutskina, and Tian (2014) show that CVCs benefit portfolio companies. Benson and Ziedonis (2010) study cases of CVC-led acquisitions. This paper contributes to the literature in two ways. First, to the best of my knowledge, it provides the first empirical exploration of why and how CVC investment decisions are made, while prior studies on CVC rationales are largely confined to surveys of managerial motives (Siegel, Siegel, and MacMillan, 1988; Macmillan, Roberts, Livada, and Wang, 2008). Second, the new evidence establishes the life-cycle pattern 6

8 of CVC investments, with emphasis on using different stages to examine CVC behaviors. This approach of exploiting the complete life cycle of CVC can be used to explore other CVC issues. 7 In broader terms, this paper builds on the literature on innovation outside firm boundaries. Nelson (1982), Telser (1982), and Jovanovic and Rob (1989) show that firms endogenously obtain innovation knowledge by searching for ideas and acquiring information externally. Aghion and Tirole (1994) theorize trade-offs that firms face when deciding to organize innovation inside or outside the boundaries of the firm. On the empirical side, Robinson (2008) shows that firms use strategic alliances to implement riskier projects when they are endowed with a set of exogenous ideas. Bena and Li (2014) show that firms with stronger innovation capabilities acquire companies with high knowledge overlaps. This paper complements that literature in two ways: first, it provides new comprehensive evidence of the under-explored CVC block in the architecture of innovation; second, it explicitly links CVC to previously studied forms of innovation efforts by tracking granular R&D, human capital, and acquisition decisions prior and subsequent to CVC investments. The remainder of the paper proceeds as follows. Section 1 describes sample construction. Sections 2 through 4 cover each stage of the CVC life cycle. Section 5 concludes. 1. Data and Measurements 1.1. The CVC Sample I construct a sample of Corporate Venture Capital units affiliated with US-based public firms, starting with the list of CVCs identified by the standard VentureXpert database. Each CVC on the 7 There is a broader business literature of CVC, see Dushnitsky (2006) and Maula (2007) for surveys. For more readings, see, e.g., Bottazzi, Da Rin, and Hellmann (2004), Dushnitsky and Lenox (2005a), Basu, Phelps, and Kotha (2011), Dimitrova (2013), Smith and Shah (2013), Ceccagnoli, Higgins, and Kang (2017), and Wadhwa, Phelps, and Kotha (2016). 7

9 list is manually matched to its unique corporate parent in Compustat by checking multiple sources including Factiva, Google, and LexisNexis. VC divisions operated by financial firms (e.g., bank affiliated or insurance company affiliated) are excluded from the sample. [Insert Table 1 Here.] The main sample consists of 381 CVC firms initiated between 1980 and Table 1 tabulates the time-series dynamic and the industry composition of CVC activities. Panel A presents the number of CVC entries and investment deals by year. Panel B summarizes the industry distribution of CVC parent firms, and industries are defined by the Fama-French 48 Industry Classification. The Business Services industry (including IT) was the most active sector in CVC investment, with 90 firms investing in 821 venture companies. Electronic Equipment firms initiated 46 CVC divisions that invested in 921 companies. Pharmaceutical firms launched 28 CVCs and invested in 254 deals. Other active sectors include Computers and Communications. In addition, I also collect information on deals conducted by CVC investors from VentureXpert. These data can help to characterize investment patterns of each investor, such as the time horizon of investment, number of companies invested, and stages of investment. They also allow us to observe the identity, final outcome, and demographic information of portfolio companies, which in turn can be used to link those entrepreneurial companies to other data sources like innovation data, as discussed below. 8 I focus on CVCs initiated no later than 2006 to allow for the whole CVC life cycle (investment behaviors, follow-up innovation, and terminations) to realize after CVC initiations. 8

10 1.2. Innovation Data Basic innovation data are obtained from the NBER Patent Data Project and from Bhaven Sampat s patent and citation data. 9 The combined database provides detailed patent-level records on more than four million patents granted by the USPTO between 1976 and It provides information on the patent assignee, the number of citations received by the patent, the technology class of the patent, and the patent s application and grant year. This database is linked to Compustat using the bridge file provided by NBER. I also link this database to startups in VentureXpert using a fuzzy matching method based on company name, basic identity information, and innovation profiles, similar to Gonzalez-Uribe (2013) and Bernstein, Giroud, and Townsend (2016). Details of the matching algorithm are explained Appendix B. 10 I employ two main variables to measure corporate innovation performance. Innovation quantity is defined as the number of ultimately successful patent applications filed by a firm in each year. A patent s year of application is used instead of the year it is granted because the former better captures the actual timing of innovation. I use the logarithm of one plus this variable, that is, ln(1 + NewPatent) (denoted as ln(newpatent)), to fix the skewness problem for better empirical properties. Innovation quality is measured as the average lifetime citations of all new patents produced by a firm in each year. Citation measures are adjusted for right-censoring as suggested by Jaffe and Trajtenberg (2002) and Lerner and Seru (2015). Similar to the logarithm transformation performed on quantity, I use ln(1 + Pat.Quality) (denoted as ln(pat.quality)). 9 For more information on the NBER Patent Data Project, please refer to Hall, Jaffe, and Trajtenberg (2001). The data used in this paper were downloaded from Sampat s data can be accessed using 10 Several Appendix tables conduct analyses on patent transactions and innovative labors. USPTO Patent Reassignment Records are used to identify patent transactions conducted by firms. The Harvard Business School inventor-level database is used to track the mobility and productivity of innovative labor around CVC activities. 9

11 Besides innovation performance, the data can also track citations made by each patent. For example, the data can observe that General Motors cites patent Internal combustion engine control for improved fuel efficiency of Tula Technology Inc. (US Patent Number , granted on August 18, 2009) in its own patent, Fuel consumption based cylinder activation and deactivation control systems and methods (US Patent Number , granted on May 17, 2016). This information helps in two ways: first, in a static term, I can identify specific underlying technologies used by each firm; in a dynamic term, this information allows me to construct variables to capture the technological diffusion among firms, especially diffusions from startups to incumbents Firm-level Measures For classic corporate governance measures, institutional shareholding information is extracted from the WRDS Thomson Reuters 13(f) data. I use total percentage institutional shareholding and the shareholding of the top five institutional investors to capture the monitoring intensity of shareholders. I also obtain the G-Index from Andrew Metrick s data library. 11 The sample is augmented with Compustat for financial statement data and with CRSP for stock market performance. The key financial variables include leverage (debt in current liabilities and long-term debt, scaled by book assets), ROA (the ratio of EBITDA to book assets), and R&D ratio (R&D expenses scaled by book assets). All variables are winsorized at the 1% and 99% levels. Key variable constructions are discussed in the Appendix. 11 Accessed using 10

12 2. The Entry of CVC To understand why incumbent industrial firms make CVC investments, I first explore the CVC entry decision, formally defined as the establishment of the CVC division. The strategic view, which mainly argues the CVCs function is to acquire innovation knowledge from the startup sector (Fast, 1978; Dushnitsky and Lenox, 2005b; Mathews, 2006), predicts CVCs to be started when external knowledge to complement internal innovation becomes more valuable than further investment in internal innovation alone. To be more specific, the theories on information acquisition and innovation model firms choosing between dedicating resources to further developing preexisting innovation and acquiring knowledge from outside that can complement and strengthen internal innovation in later periods (Nelson, 1982; Telser, 1982; Jovanovic and Rob, 1989). Resources dedicated to information acquisition, such as through CVC, are determined by the quantity and quality of existing ideas available to the firm the smaller (lower) the quantity (quality) of existing innovation ideas becomes, the more likely the firm will implement CVC in search of better innovation paths. Accordingly, CVCs are more likely to be launched following innovation deteriorations. The root of the agency view of CVC is the long-lasting literature in corporate finance showing that managers extract private utility by expanding firm boundaries, which in turn affects their decisions on investments (Jensen, 1986; Stulz, 1990) and on the diversification of the corporation (Denis, Denis, and Sarin, 1997). Proponents of this view would argue that CVCs manifest managers desire to enjoy managerial perks via venture investing or to build an empire, rather than to create value for the firm. Accordingly, CVCs tend to form in firms whose shareholders are unable to discipline managers. The financial view, which builds on the venture investor nature of CVCs, suggests that CVCs sim- 11

13 ply reflect incumbent firms motivation to earn financial returns from the promising entrepreneurial sector. This view has ambiguous predictions regarding when a firm is more likely to enter CVC. On the one hand, it argues that firms can shift investment focus from internal R&D to external investment such as CVC when internal innovation opportunities are poor, following the classic Q-theory argument (Hayashi, 1982). On the other hand, this view can predict firms entering the CVC business following a period in which their industry knowledge becomes more advantageous in assessing venture opportunities, like when their internal operation prospers Baseline Model Specification The baseline model examines the CVC entry decision on the firm-year panel of US public firms with valid ROA, size (logarithm of total assets), leverage, R&D ratio, and at least $10 million in book assets. Only innovative firms, defined as those that filed at least one patent application that was eventually granted by the USPTO, are included. Industries (3-digit SIC level) with no CVC activities during the whole sample period are excluded. The empirical model takes the following form: I(CVC) i,t = α industry t + β I τ Innovation i,t 1 + β G Governance i,t 1 + γ X i,t 1 + ε i,t, (1) where I(CVC) i,t is equal to one if firm i launches a CVC unit in year t, and zero otherwise. 12 τ Innovation i,t 1 is the change of firm innovation performance over the past τ years ending in t 1, which naturally differences out firm-specific innovation levels. I use a three-year (τ = 3) innovation shock throughout the main analysis and report robustness checks using other horizons in 12 Since the model predicts CVC launches, a CVC parent firm naturally drops out of the sample after the initiation. It re-enters after one CVC life cycle concludes. 12

14 the Appendix. Governance measures include Institutional Shareholding and the G-Index. Firm-level controls X i,t 1 include ROA, size, leverage, and R&D ratio. Industry-by-year fixed effects are included to absorb industry-specific time trends, and industries are defined by the Fama-French 48 Industry Classification. [Insert Table 2 Here.] Table 2 presents descriptive statistics of the regression sample. I show for both firm-year observations when a CVC division was initiated and those observations when a CVC was not initiated. CVC parents are typically large firms. On average, a CVC parent has $10.1 billion in book assets (median is $2.4 billion) just before launching its CVC unit, whereas non-cvc parent firms have less than $3 billion in book assets (median is $0.2 billion). CVC parent firms are innovation intensive in terms of patenting quantity, echoing the size effect. CVC parent firms experience more negative innovation shocks before starting their CVC divisions they on average experience a -7% (-10%) change in patenting quantity (quality) within the three years prior to launching their CVC units, compared to the control firms, which experience a 12% (8%) shock. Corporate governance variables, G-Index and Institutional Shareholding, are comparable between the two subsamples Baseline Regression Results Table 3 presents the Ordinary Least Squares (OLS) estimation of a linear probability model (1). Column (1) focuses on the effect of changes in innovation quantity. The coefficient of is negative and significant, meaning that a more severe decline in innovation quantity in the past three years is associated with a higher probability of initiating CVC investments. This estimate translates a one-standard-deviation decrease (σ-change) in ln(newpatent) into a 26% increase 13

15 from the unconditional probability of launching CVC unites. Column (2) studies the effect of deterioration in innovation quality. The coefficient of means that a one-standard-deviation decrease in ln(pat.quality) increases the probability of CVC initiation by 34%, and this is economically comparable to that in column (1). Column (3) simultaneously estimates the effects of changes in innovation quantity and quality. The estimates are largely unchanged compared to columns (1) to (2). Overall, CVC entries typically follow deteriorations in internal innovation of a firm. [Insert Table 3 Here.] Columns (4) to (6) study the effects of classic corporate governance measures on CVC entry. Neither institutional shareholding nor G-Index has any real influence on the CVC entry decision. In column (4), I use total institutional shareholding to measure governance intensity and find it has a positive, insignificant effect on CVC entry. The result is similar when we use the shareholding of only the top five institutional shareholders. In column (5) we focus on G-Index (which unfortunately restricts the sample size significantly). G-Index also does not have explanatory power on the initiation of CVCs. It is worth stressing the importance of incorporating industry-by-year fixed effects in model estimations. Previous studies highlight the possibility that certain industry-specific technology shocks could be driving innovation changes and organizational activities at the same time (Mitchell and Mulherin, 1996; Harford, 2005; Rhodes-Kropf, Robinson, and Viswanathan, 2005). After absorbing this variation using industry-by-year fixed effects, the results in Table 3 are identified using the cross-sectional variation in innovation dynamics within an industry-by-year cohort. I conduct an array of robustness checks to confirm that the CVC initiation results are not driven 14

16 by the sampling process or specifications. In Table A1, I report the analysis using alternative horizon parameters τ. In Table A2, I estimate the probability of CVC entry using a hazard model developed by Meyer (1990) and utilized in Whited (2006), which fits this paper s context due to its capability of incorporating time-varying predictors and stratified groups. I find similar results in those analyses. I also show that the results are robust to removing firms that are from specific industries or that are located in specific locations (Table A3) Assessing the Agency View of CVC Entry Table 3 provides supporting evidence for the strategic view and the financial view of CVC, but it is largely inconsistent with the agency view. However, to cleanly interpret why CVC entries follow internal innovation deteriorations, it is necessary to understand the variations that drive innovation changes in the first place. For example, it could still be the case that Table 3 signifies that an entrenched manager hinders innovation and simultaneously leads to the initiation of CVC as a pet project. As a result, to more confidently rule out the agency interpretation of CVC entry, I need an exogenous shifter that could affect an individual firm s ability to generate innovation ideas internally (the first stage), but which is unlikely to relate to CVC investments through the agency channel (the exclusion restriction). The main idea of the empirical strategy is to exploit the influence of exogenous technological evolution on firm-specific innovation knowledge. In other words, the instrument variable will shock the individual firm s ability to generate innovation using exogenous changes to the usefulness of its accumulated knowledge. For example, the empirical strategy will exploit the cases in which a firm specializing in 14-inch hard disk drives (HHDs) becomes less able to innovate when the technology 15

17 moves on to 8-inch HDDs. 13 To implement the idea of measuring the influence of exogenous technological evolution on an individual firm s capability to innovate, I build on the literature of bibliometrics and scientometrics, which measures the obsolescence and aging of a scientific discipline 14 using the dynamics of citations referring to the specific field. In particular, I construct a firm-year level variable, termed as Knowledge Obsolescence (Obsolescence for short), to capture the τ-year (between t τ and t) rate of obsolescence of the knowledge possessed by a firm as of t τ. For each firm i in year t, this instrument is constructed in three steps, formally defined in formula (2). First, firm i s predetermined knowledge space in year t τ is defined as all the patents cited by firm i (but not belonging to i) up to year t τ. This fixed set of patents proxies for the underlying technological knowledge that firm i managed to accumulate up to t τ. I then calculate the number of external citations (made by firms other than i itself) received by this fixed KnowledgeSpace i,t τ in t τ and in t, respectively. Last, Obsolescence τ i,t is defined as the rate of change between the two, which naturally absorbs effects of the size of the firm and its knowledge space. Formally, Obsolescence τ i,t = [ln(cit t (KnowledgeSpace i,t τ )) ln(cit t τ (KnowledgeSpace i,t τ ))]. (2) A larger Obsolescence means a greater decline of the value and utility of a firm s knowledge within the τ-year period, as captured by the fact that fewer new patents build on that knowledge base. 13 Indeed, new technologies come and go, taking generations of companies with them (Igami, 2017). See also Christiansen (1997). 14 The methodology has been similarly applied to evaluate the impact of specific technologies and individual research, among others. 16

18 Knowledge Obsolescence and Innovation. The idea that knowledge obsolescence affects innovation builds upon two theoretical pillars. First, the knowledge stock of an individual or institution determines the quantity and quality of its innovation production. Jones (2009) shows that a negative shock to the value of a firm s accumulated knowledge space implies a longer distance to the knowledge frontier and a higher knowledge burden to identify valuable ideas and produce radical innovation. Bloom, Schankerman, and Van Reenen (2013) show that firms working in a fading area benefit less from knowledge spillover, which in turn dampens growth in innovation and productivity. Second, knowledge itself ages. In the past few decades, several disciplines have developed the concept of the obsolescence of knowledge, skills, and technology. The most famous result might be, roughly speaking, that half of our knowledge today will be of little value (or even proven wrong) after a certain amount of time (i.e., half-life), and this half-life is becoming shorter and shorter (Machlup, 1962). Economists have studied the effect of obsolescence of knowledge and skills on labor and industrial organization, as well as on aggregate growth (Rosen, 1975). Empirically, the effect of knowledge obsolescence on corporate innovation is validated in the first-stage regression, in which I instrument τ Innovation i,t with Obsolescence τ i,t using the following form: τ Innovation i,t = π 0,industry t + π 1 Obsolescence τ i,t + π 2 X i,t + η i,t. (3) [Insert Table 4 Here.] Table 4 columns (1) and (3) report results where Innovation is measured using the quantity and quality of new patents, respectively. Results show that a faster rate of Knowledge Obsolescence is associated with weaker internal production of innovation. The estimate of in column (1) 17

19 translates a 10% increase in the rate of obsolescence of a firm s knowledge space into a 1.14% decrease in its patent applications; this same change is associated with a 1.28% decrease of its patent quality. The F-statistics of these first-stage regressions are both well above the conventional threshold for weak instruments (Stock and Yogo, 2005) Exploiting the Obsolescence-driven Variations. The first-stage regression (3) allows us to extract variations to innovation driven by plausibly exogenous trends of knowledge obsolescence and that are independent to agency frictions. The fitted value from this model, denoted as is then used in the second-stage regression, Innovation, I(CVC) i,t = α industry t + β τ Innovation i,t 1 + γ X i,t 1 + ε i,t, (4) and columns (2) and (4) of Table 4 show the estimation results. The effect of obsolescencedriven innovation shocks Innovation on starting a CVC unit is both economically and statistically significant. The coefficient of in column (2) translates a σ-change in ln(newpatent) to a 26% change in the probability of launching CVC investment. The gaps between the OLS estimates (in Table 3) and the 2SLS estimates are small. This comparison suggests that the agency-related interpretation does not seem to drive the OLS estimation initially. In a reduced form, Table 2 also reports summary statistics for Obsolescence. The number of citations received by a firm s predetermined knowledge space decays by 8% in the control group, which can be interpreted as a benchmark three-year natural decay of knowledge. Firms knowledge spaces on average decay by 29% in the three years before initiating a CVC division, which demonstrates a much more severe hit by the technological evolution. Table 4, column (5) 18

20 reports a reduced-form regression in which Obsolescence is used to explain the decision to launch a CVC program. The positive coefficient indicates that firms experiencing larger technological decays are more likely to start CVC activities Discussions on the Empirical Assumptions. To be clear, the motivation behind the Obsolescence instrument and related analysis is to further rule out the concern that the agency view could still be at play in Table 3. To further justify Obsolescence to be a valid source of exogenous variation to innovation that does not affect CVC investments through the agency channel, I provide additional discussions on this assumption in this section. The first building block of the instrument is the formation of the KnowledgeSpace, defined as the set of patents that a firm cites in its previous patents. One potential concern is that a firm s knowledge space can signal the capability of its manager, which in turn can affect its innovation policy. I assess this concern both qualitatively and quantitatively. On the one hand, historical poor management is unlikely to affect the specific timing of current CVC launches in other words, it is unlikely that poor innovation decisions before t τ should lead to CVC investments in t. On the other hand, in an additional analysis, I construct the pre-determined knowledge space of firm i based on its citations made before t 10 and track the obsolescence of this knowledge space from t 3 to t. 15 The possibility that the managerial vision ten years ago still strongly affects CVC decisions today is thin, thus better disentangling firms knowledge spaces from concurrent managerial decisions. Table 5, Panel A presents results that are qualitatively and quantitatively similar to Table 4. [Insert Table 5 Here.] 15 This analysis necessarily focuses on the sample in the later period and firms that have longer patenting histories. 19

21 The second key component of the instrument is the citation dynamics regarding knowledge spaces. One might worry that the firm itself could be a main driver of the technological evolution. For example, a manager might decide to change the course of innovation areas using CVC, and this change could potentially lead to citation changes to the firm s own knowledge base (say, a diesel engine maker enters the gas engine industry and stops citing diesel engine technologies). To be on the conservative side, I have excluded patents owned by the firm from its own knowledge space and all citations made by the firm itself (self-cites) in the construction of Obsolescence. In other words, any direct impact of a firm itself on the citation dynamic is eliminated from the measure. In addition, I conduct an empirical test in which I repeat the 2SLS analysis in subsamples of firms with high vs. low innovation impact, where innovation impact is categorized using the median of the number of patents possessed by the firm in each year. The idea is that those low-impact firms are less likely to endogenize the technological evolution. I report the results in Panel B of Table 5, and the results are both qualitatively and quantitatively similar to Table 4. The results are also robust when defining innovation impact using total patent applications in the past three years or using market valuation. Overall, Table 5 suggests that despite potential concerns, the relation between innovation deterioration and CVC launches does not seem to be driven by the variation in agency frictions. These results, combined with the evidence that both institutional shareholding and the G-Index both lack power in explaining CVC entry, lend little support to the agency view of CVC Assessing the Financial View of CVC Entry What is left unclear is whether the CVC entry in response to innovation deteriorations is motivated by the desire to strategically acquire innovation knowledge or to seek pure financial 16 A more detailed discussion on the Obsolescence variable is provided in Appendix C. 20

22 returns. As will be described below, either motivation is plausible. As a result, instead of attempting to rule out either of the interpretations, the goal here at the entry analysis is milder. In specific, I conduct two new analyses hoping to assess to what extent we can distinguish two views at entry. Additional analyses and discussions will be provided in later stages of the life cycle. The first analysis examines whether CVC entries are accompanied by declines of internal R&D. If CVCs reflect corporate actions to seek higher financial returns when internal investment opportunities dry up, we would expect an internal R&D decrease together with the patenting decrease to reflect the shift away from internal investment. In contrast, if CVCs are for strategic complementarities, one could expect R&D to be stable and the innovation direction to be shifted toward the technologies of portfolio startups. In Table 6 Panel A, I show that measures of innovation input (i.e., R&D) expenditures scaled by total assets or sales do not affect the CVC entry decision. Putting this result into the context of Table 3, the interpretation is that CVC is not a way for firms to shift from internal innovation to external innovation, but for them to respond to deteriorating innovative capabilities. The second analysis examines whether innovation deteriorations motivate financial or strategic CVCs. I categorize CVCs in the sample into financial- or strategic-driven by collecting information disclosed at the announcement of CVC initiations using a news search, following a similar approach as Dushnitsky and Lenox (2006). For each CVC in the sample, I search for media coverage and corporate news at its initiation using LexisNexis, Factiva, and Google. Based on this compiled information, CVCs are coded as financial and strategic. When the main objective of a CVC unit is difficult to categorize, I code it as unknown. In the end, I successfully categorized 204 CVCs. The logic behind the analysis is straightforward: if financial return is a key driver behind the relation, the results in Table 3 and Table 4 should hold at least as strongly when focusing on the 21

23 initiations of the small set of financial CVCs. I report the results in Table 6, Panel A, which shows that innovation deteriorations motivate strategic CVCs with much higher intensities, suggesting that the main effect that innovation deteriorations have on CVC decisions is mostly driven by strategic considerations. Meanwhile, financial CVCs are less responsive to internal innovation performance. [Insert Table 6 Here.] Admittedly, it is difficult to dispute that financial returns are important for any corporate investment; in fact, a small set of CVCs declare themselves as financial return-driven. However, the additional evidence provided in Table 6 suggests that the strategic view of CVC is the main driving force behind CVC entries. 3. The Investment of CVC To further examine the strategic view and the financial view of CVC, the analysis moves on to the investment stage of the CVC life cycle. Under the strategic view, CVCs are adopted to help parent firms learn new innovation knowledge from the entrepreneurial sector and then to further implement those new technologies to complement their internal innovation. Accordingly, CVCs are expected to invest in startups that can provide newer and more useful knowledge and to integrate this new technological information into parent firms organic R&D. Under the financial view, in contrast, CVCs are investment vehicles for incumbent firms to exploit their industry knowledge in selecting and monitoring targets and to harvest financial returns. Accordingly, CVCs are expected to act like financial return-driven IVCs and to invest in companies about which they possess advantageous information and that are easier for them to monitor. 22

24 3.1. CVC Portfolio Formation I start by examining characteristics that lead to the formation of a CVC-startup deal, and the key test is an empirical matching model between CVCs and portfolio companies. I first build a data set of all potential CVC-startup pairs by pairing each CVC i with each entrepreneurial company j that had ever received an investment by a VC. I remove such pairs when the active investment years of the CVC firm i (between initiation and termination) and the active financing years of company j (between the first and the last round of VC financing) do not overlap. For each CVC-startup pair i- j, I first construct two variables, Technological Proximity (Tech- Proximity) and Knowledge Overlap (Overlap), to assess the role of technological distances on CVC-startup matching. TechProximity is calculated as the cosine similarity between the CVC s and the startup s vectors of patent weights across different technology classes (Jaffe, 1986; Bena and Li, 2014). A higher TechProximity indicates that the pair of firms works in closer areas in the technological space. Overlap is calculated as the ratio of (1) numerator: the number of patents that receive at least one citation from CVC firm i and one citation from entrepreneurial company j; to (2) denominator: the number of patents that receive at least one citation from either CVC i or company j (or both). A higher Overlap means that the pair of firms shares broader common knowledge in their innovation. 17 In addition to measuring the technological distance for each pair, I also construct two measures to capture the geographic distance. Local is a dummy variable indicating whether CVC firm i and company j are located in the same Commuting Zone (CZ). CZ is used as the main geographic 17 Both Technological Proximity and Knowledge Overlap are measured as of the last year before CVC i and company j both enter the VC-startup community. For example, if firm i initiates the CVC in 1995 but company j obtained its first round of financing in 1998, the measure is constructed using the patent profiles in The rationale for this criterion is to mitigate the potential interactions between CVCs and startups before investment, thus providing a clean interpretation of the estimation. 23

25 delineation because it has been shown to be more relevant for geographic economic activities (Autor, Dorn, and Hanson, 2013; Adelino, Ma, and Robinson, 2017). I also include the natural logarithm of the distance between firm i and startup j (accurate at ZIP Code-level, kilometers). The empirical test exploits a reduced-form matching model on this sample of CVC-startup pairs to predict the decision of CVC i investing in company j, in the following form, I(CVC i -Target j ) = α + β 1 TechProximity i j + β 2 KnowledgeOverlap i j + β 3 Local i j + β 4 Distance i j + γ X i, j + ε i, j, (5) where the dependent variable I(CVC i -Target j ) indicates whether CVC i actually invests in company j (i.e., the realized pair). In X i, j I control for CVC (i)-level characteristics including number of annual patent applications and average citations of patents; I also control for those innovation characteristics at the startup ( j)-level. [Insert Table 7 Here.] Table 7 presents coefficients estimated from model (5). In column (1), a positive and significant coefficient means that the Technological Proximity between a CVC and an entrepreneurial company increases the likelihood of CVC deal formation. This means that a one-standard-deviation increase of TechProximity between a CVC parent firm and a startup doubles the probability that an investment relationship is formed. Column (2) examines Knowledge Overlap. The negative coefficient means that after conditioning on technological proximity, CVC parent firms prefer to invest in companies about which they have more limited knowledge. A one-standard-deviation increase of Overlap leads to a 40% decrease in investment probability. In column (3), I explore whether CVCs are more or less likely to invest in geographically 24

26 proximate firms. The venture capital literature, and the investment literature more broadly, has documented a home (local) bias phenomenon when investing in companies that are geographically closer, financial return-driven investors can better resolve the information asymmetry problem and conduct more efficient monitoring (Cumming and Dai, 2010; Hochberg and Rauh, 2012; Bernstein, Giroud, and Townsend, 2016). In columns (3) and (4), however, I find that CVCs do not really invest in their home companies with and without controlling for the distance measure. The dummy variable indicating that the CVC and the startup are located in the same Commuting Zone negatively affects the probability of investment. This finding is consistent with the strategic explanation that CVC parent firms can acquire innovation knowledge from startups in the same CZ through local innovation spillover (Jaffe, Trajtenberg, and Henderson, 1993; Peri, 2005; Matray, 2016), which decreases the marginal benefit of making a CVC investment in them. Overall, Table 7 shows that CVC investment behaviors differ greatly from the well-studied IVCs. CVCs invest in companies that possess knowledge complementary to the parent firm, rather than those of which they possess advantageous information. In addition, they invest in companies at longer distances at the expense of monitoring difficulties in order to acquire knowledge that can otherwise be hard to obtain. Those all appear to be more consistent with the strategic view of CVCs Identifying Integrated Strategic Complementarities Do parent firms integrate CVC-led innovation knowledge to complement their internal innovation? This section tests whether the investment relationship between a CVC and an entrepreneurial venture leads the parent firm to adopt innovation produced by the entrepreneurial venture, which 18 Appendix D provides a more in-depth discussion on the investment patterns of CVCs and their differences with traditional IVCs. 25